Compact Autoencoder Architectures for Anomaly Detection in Electric Drivetrain Current Sensors: a Comparative Study

Accurate current sensing in the drivetrain of electric vehicles is critical for safe and efficient operation. Errors in the current sensor can lead to significant consequences, such as inaccuracies in the torque estimation and deviations from ideal operating points. This can effect the overall efficiency of the drivetrain. Robust anomaly detection techniques can address these challenges. In this study four different types of autoencoder architectures: Long Short-Term Memory (LSTM)-based, LSTMVariational Autoencoder (LSTM-VAE), transformer-based, and a fully convolutional network are investigated. Using the knowledge from a healthy system, the examined models are trained using unsupervised learning techniques to capture normal operating patterns. A key focus lies in the potential to employ in embedded devices, necessitating the development of highly compact and resource-efficient neural networks. Through comparative analysis, an ultra-compact LSTM architecture, the Tiny-LSTM, is identified as the most suitable model for the given task, offering significant advantages in resource efficiency while showing the best performance.